Am Montag, dem 18.03.2024 um 09:33 +0800 schrieb Yu Kuai:

you might need to learn some tools like blktrace or bpftrace to find out
which thread is issuing IO to sdb1.

Thnaks for the hint, I'll play around with these tools.

Some other musings: as this is a RAID-1 array, and both sda1 and sdb1 are "identical" (both are
flagged with write-mostly), I *should* see identical write patterns to sda1 and sdb1?

If we look at my iostat output from above:

Device             tps    kB_read/s    kB_wrtn/s    kB_dscd/s    kB_read    kB_wrtn    kB_dscd
md0               2,20         0,00         8,80         0,00          0         44          0
nvme0n1p3         3,60         0,00         9,50         0,00          0         47          0
sda1              3,80         0,00         9,50         0,00          0         47          0
sdb1             54,20         0,00     26223,10         0,00          0     131115          0
 

44 kb have been written to md0, the md subsystem converts these to writes to the RAID members (plus
some overhead like bitmaps and stuff)

The 47 kb written to nvme and sda1 is what I'd expect to see. But the 130 MB to sdb1 are wrong...

btw, when I run this tests on kernel 6.1.76, I get identical writes to all RAID members:

Device             tps    kB_read/s    kB_wrtn/s    kB_dscd/s    kB_read    kB_wrtn    kB_dscd
md0               5,40         0,00        67,20         0,00          0        336          0
nvme0n1p3         3,40         0,00        68,00         0,00          0        340          0
sda1              3,40         0,00        68,00         0,00          0        340          0
sdb1              3,40         0,00        68,00         0,00          0        340          0


Wild guess: the (external) USB device sdb1 is using a huge "transfer size", so when only a few
sectors are written to sda1, megabytes are written to sdb1?

How could I proove this?

thanks, Michael

-- 
Michael Reinelt [off-list ref]
Ringsiedlung 75
A-8111 Gratwein-Straßengel
+43 676 3079941